# Changelog

All notable changes to this project will be documented in this file.

The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).

## [Unreleased]

### Fixed

- **A failed publish was never retried, and the watchdog could not see a stale feed** (alpha-engine-config-I12901, audit 2026-10-30 PROD-05). `save_metadata` wrote `audio_file` before `publish_to_s3` ran, and the front-door dedup treated any metadata with audio as a finished episode. A publish that failed once was therefore skipped by every later run. Episode metadata now records generation (`published_at: null`) separately from delivery. The delivery checkpoint is set only after `verify_publication` confirms that the published `feed.xml` carries an item for that date and edition and that its enclosure object exists and is non-empty. An episode that was generated but never verified resumes publication from its existing MP3 on the next ordinary run, with no script or TTS call. That covers failures at MP3 upload, at feed upload and at verification. `--no-publish`, `--script-only` and `--publish-only` keep their meaning: `--publish-only` also checkpoints a pending episode once verified. Metadata writes are atomic. Metadata written before this change, with no `published_at` key, counts as published. The watchdog now also runs `check_feed_delivers`, so a fresh MP3 behind a missing, malformed or old feed exits 1.

- **Episode scripts had no length floor** (alpha-engine-config-I11831). A
  near-empty script that passed the tool-call, grounding and topic guards went
  straight to TTS/publish and was announced with the same check mark as a full
  episode (2026-09-21: ~200 words aired as 53 seconds). `generate_script` now
  hard-aborts an attempt whose script is under `min_script_words` (default 500,
  per-edition mapping supported, `0` disables), so the fallback tier runs
  instead; the success ping now reports the floor and replaces the check mark
  with a SHORT warning when a published script is under it.
- **The README's `coverage-96%` badge was hand-set and wrong** — the suite
  measures 92.54%. A badge whose value a human types renders identically to a
  measured one, so a reader cannot tell them apart, and it becomes false the
  moment reality moves without anyone editing a file
  (repository-baseline-policy.md §5.1). The badge now renders a shields.io
  `endpoint` document that `scripts/publish_coverage_badge.sh` rewrites from
  the same `.coverage` file the gate reads, on every push to `main`. The
  `python` and `License` badges likewise now come from PyPI metadata and the
  repo's own LICENSE rather than from typed literals.
- **The coverage floor was 80% against 92.54% measured**, so 12 points of
  coverage could be lost without the gate noticing. Raised to 92 as a ratchet,
  and `tests/test_coverage_scope.py` now asserts the measurement *scope* —
  the way a coverage gate stops being honest is by narrowing what it measures
  rather than by lowering the number, which reads as an improvement in every
  report (§4.2 C5).

- **Every AM episode since the 2026-08-13 litellm_proxy router migration
  aired unexecuted tool-call XML instead of a script.** Every prompt
  (`prompt.md`/`prompt_weekend.md`/`prompt_public.md`) unconditionally
  instructs the model to call a `web_search` tool. On a transport with no
  server-side search (litellm, openai, direct), `call_with_grounding_degrade`
  degrades to a bare `complete()` call with no tools attached — but the
  system prompt still told the model it had one. The model dutifully
  hallucinated `<tool_calls>`/`<invoke name="...">` XML as its entire
  response, and nothing caught it: the content-grounding guards that would
  normally reject a hollow script are the exact guards this path skips
  (zero searches/citations is *expected* here by construction). Root-caused
  live 2026-08-19 against the production box (`i-09b539c844515d549`) —
  `script_chars` in every `{date}-am.llm_decision.json` since 2026-08-13 was
  198-911 (normal: 3,000-15,000). Fixed at two layers: (1) the `complete()`
  degrade path now explicitly tells the model it has no tools and to use
  only the provided news context; (2) a backstop guard rejects any script
  containing unexecuted tool-call XML and raises, engaging the same
  fallback-tier cascade as any other hard-abort, rather than publishing it.
- **The unconfigured-`llm` default no longer silently reaches for a
  hardcoded Anthropic model on a Nous Ergon (`MORNING_SIGNAL_USE_SSM=1`)
  deployment.** Production's SSM config always sets `llm`; reaching the
  default with SSM mode active now raises `LLMConfigError` (a deploy
  defect) instead of defaulting. Self-hosted installs (no SSM, no router)
  are unaffected — they keep the pre-migration literal default, since
  those installs have no registry to derive a route from.

## [0.2.2] — 2026-08-17

### Fixed

- **The router path no longer silently shadows the krepis registry's
  output-token budget.** `_resolve_router_group` passed
  `spec.max_tokens or config.get("max_tokens", 4096)` to
  `resolve_group_spec` — `spec` is an inert `ModelSpec` whose `max_tokens`
  krepis' dataclass forces to always be a truthy int, so the `or` never
  once fell through, and `resolve_group_spec` never once deferred to the
  registry's row for whichever model the `llm` group resolves to. Same bug
  as crucible-evaluator's Director (alpha-engine-config-I6396): a call-site
  literal silently wins over the registry-owned budget
  (model-router-policy §2), with no warning and no trace in any log. The
  krepis `high` group currently resolves to DeepSeek V4 Pro Max, a
  reasoning model — `max_tokens` bounds reasoning AND content from one
  shared pool (alpha-engine-config-I6901), so a budget sized for a
  non-reasoning answer can starve content generation entirely. `max_tokens`
  in `config.yaml` is now an explicit opt-in override on the router path;
  leaving it unset lets krepis use the registry's row. The direct-Anthropic
  self-hosted default path is unaffected (no registry to defer to there).

## [0.2.1] — 2026-08-02

### Changed

- **Primary LLM can be a krepis router model group (`litellm:high` etc.).**
  Transports without server-side web search (litellm, plain openai) fall
  through from `complete_grounded` to `complete()`, grounding on the
  pre-fetched `news_context` digest instead of live search. Citation and
  `required_search_topics` guards are skipped on that path rather than
  falsely aborting an empty-search result. Production points
  `/morning-signal/config-yaml` at `{"provider": "litellm", "model": "high"}`
  so the local LiteLLM proxy owns primary routing + its own fallback chain.
- **Cascade catches all exceptions, not just `RuntimeError`.** A primary
  billing failure (OpenRouter HTTP 402 on 2026-08-02) is an
  `APIStatusError`, not a `RuntimeError` — the previous `except RuntimeError`
  let it kill the process before the Anthropic ultimate tier engaged.
  Cascade now catches `Exception` at every tier so non-RuntimeError provider
  failures still fall through.

### Added

- **`schedule:` config block — per-date scheduled content overrides
  (OPTIONAL ADD-ON, default OFF).** When enabled, a schedule manifest JSON
  (`schedule/schedule.json` in the podcast bucket by default; schema in
  `docs/schedule-schema.json`, `schema_version` 1) is read from S3 at
  generation time. An entry for the run date can: `override` — devote the
  whole episode to a scheduled deep-dive topic instead of regular
  programming; `extend` — add one extra segment on top of the regular
  lineup; or `skip` — produce no episode at all that day (the
  console-editable counterpart of `skip_dates:` below; both sources are
  honored by the generate guard AND the watchdog, and `generate --force`
  overrides either). Deep dives are guaranteed live-researched, not
  written from memory: the entry's topic becomes a per-run
  `required_search_topics` guard riding the existing coverage +
  forced-`web_search` recovery machinery (in `override` mode it REPLACES
  the config-declared topics — segments that intentionally don't air must
  not trigger recovery or a fatal abort). Basic form unaffected: with the
  block absent/disabled there is ZERO behavior change (no S3 read, no new
  deps, no new CLI surface); even when enabled a missing manifest is a
  silent no-op, and any schedule READ failure fails soft to the regular
  episode with a WARN log + Telegram alert (`watchdog.send_alert`). After
  a schedule entry is applied, a best-effort marker is written to
  `schedule/applied/{date}-{edition}.json` so the (separate) console
  schedule page can show the entry actually aired. New module:
  `morning_signal.schedule_override`.
- **Custom-episode RSS titles.** A schedule-override `override`-mode
  episode (regular programming replaced entirely by a scheduled deep-dive
  topic) now gets its own feed item title, derived directly from the
  entry's `topic` field — e.g. `Morning Signal — NVIDIA earnings deep
  dive` instead of the generic `Morning Signal — 2026-07-14 AM`. Regular
  programming and `extend`-mode episodes (which still air the normal
  lineup plus one extra segment) are unaffected and keep the default
  date-stamped title. New `episode._derive_title`; `episode.save_metadata`
  gains an optional `title` field (additive — omitted/`null` for existing
  episode JSON, so old metadata keeps working unchanged).
- **`skip_dates:` config list — skip specific dates entirely.** Days the
  listener can't listen (travel, vacation) as ISO `YYYY-MM-DD` strings: the
  generate run no-ops cleanly for BOTH editions (exit 0, no failure email,
  no Claude/TTS/S3 spend), and the freshness watchdog treats the absent
  episode as expected instead of paging. Distinct from the NYSE trading
  calendar, which only reshapes editions (weekends/holidays still ship a
  weekend AM). `generate --force` overrides a skip date for a one-off manual
  run; `--publish-only` (feed rebuild) is unaffected. Malformed entries fail
  loud at config parse (`morning_signal.config.parse_skip_dates`) rather than
  silently never matching a run date.
- **`required_search_topics` entries can be scoped to editions (`editions`).**
  Different editions run different prompts with different segments — e.g. a
  weekday edition with a political-pulse segment and a `weekend` edition (the
  non-trading-day prompt) with no politics at all. A topic carrying an
  `editions` list (values: `am`, `pm`, `weekend`) is enforced only for those
  editions, so a weekday-only topic no longer falsely aborts the weekend pod
  that legitimately never searches it. Omit `editions` to enforce on every
  edition (unchanged default). `unmet_required_topics` gains an optional
  `edition` arg; `generate_script` passes `"weekend"` on non-trading days.

- **Per-segment search guard (`required_search_topics`, default none).** The
  global `min_web_searches` floor only asserts an edition was grounded
  *somewhere* — it does not guarantee a *specific* search-critical segment was
  covered. With a tight `web_search_max_uses` budget the model can spend its
  searches on earlier segments and reach a no-digest segment (e.g. a political
  pulse sourced only from Truth Social / X) with no budget left, then write it
  from memory while still clearing the floor. `required_search_topics` lets the
  operator assert, per topic, that at least `min_matches` (default 1) searches
  actually targeted it (case-insensitive keyword match on the query); an unmet
  topic ABORTS before publish. Default empty = no-op (OSS-safe); topics are
  declared in the operator's config alongside the prompt that defines those
  segments. New `search_telemetry.unmet_required_topics` + public
  `search_telemetry.extract_searches`. (2026-06-17: political segments were
  silently dropped two days running because the live `web_search_max_uses` was
  throttled to 5 — far below what 9 segments, 4 of them political, needed.)

- **Pre-fetched news digest is a hard prerequisite when enabled
  (`news_context.required`, default `true`).** A soft-failed digest
  (missing / malformed / **stale** — its `date` != the episode's date /
  empty) now makes `load_news_context` **raise** and aborts the pod
  before publish, instead of silently degrading into an episode narrated
  without the news it was meant to carry. The freshness watchdog then
  catches the absent episode. Set `news_context.required: false` to opt
  back into the original fail-soft (warn + web-search-only) behavior.
  `generate_script` passes the episode date so staleness can be checked.

- **Fail-loud web-search floor guard.** `generate_script` now aborts
  before any TTS/publish when an edition runs fewer than
  `min_web_searches` web searches (default `1`). A zero/low-search
  edition is almost always model-confabulated rather than grounded in
  live news; raising lets the silent-failure watchdog catch the absent
  fresh episode instead of a hallucinated one going live. OSS users with
  a prompt that legitimately needs no live search can set
  `min_web_searches: 0` to opt out.

### Changed

- **Pre-fetched news context reframed as supplementary.** The injected
  news block previously instructed the model to "use the items below …
  do NOT web-search them", which made the model skip web search entirely
  — including segments the digest never covers (e.g. the political /
  Truth Social pulse) — and hallucinate a full episode with
  `web_search_requests == 0` (2026-06-16 incident). The block is now
  framed as a supplementary starting reference: the model is told it MUST
  still web-search every segment, that the digest omits the political
  segments, and to prefer fresh search results over a pre-fetched item on
  conflict.

### Removed

- **Public-topics mode and segmented generation.** Dropped the
  `public_topics_mode` / `generation_mode: segmented` machinery — the
  `prompt_public.md` 10-topic catalog, the `topic_rotation.py` wildcard
  rotation, per-topic `generate_segments` / `generate_freeform_segment`
  generation, `tts.synthesize_segments`, and the `enforce_char_budget`
  circuit-breaker. These existed to dogfood a hosted multi-tenant
  "topic catalog" product that is not part of this open-source project.
  The engine now has a single, fully prompt-driven generation path:
  customize `prompt.md` (and `prompt_weekend.md`) to define your episode's
  content and format — see the `src/morning_signal/data/prompt-*` starters.
  Removed config keys (now ignored if present): `public_topics_mode`,
  `generation_mode`, `segment_search_max_uses`, `segment_word_target`,
  `episode_max_chars`, `freeform_topic`, `freeform_max_chars`.

## [0.1.3] — 2026-06-10

### Added

- **`morning-signal watchdog` — outcome-based freshness check.** A new
  subcommand that verifies today's episode actually landed in S3 (present +
  fresh within `--max-age-hours`, default 6) and exits non-zero otherwise.
  Run it on a timer shortly after the generate slot to catch the SILENT
  failure modes the in-process Telegram notifier cannot report: a
  bootstrap/AssumeRole/SSM failure (which happens *before* the notifier creds
  are loaded), the timer never firing, or an OOM kill. `--notify` sends a
  Telegram alert via the configured notifier for self-hosters. Example systemd
  units ship under `infrastructure/`.

### Fixed

- **S3 client now uses the bucket's region during SSM bootstrap.** `aws.py`
  created the prompt-fetch S3 client in the SSM region (us-east-1) instead of
  the bucket's `s3_region` (us-west-2), adding a cross-region redirect
  round-trip per boot. (ROADMAP S2.)
- **Legacy `generate_episode.py` shim no longer exits 2 on a bare
  invocation.** `_is_legacy_invocation` now routes a no-arg call to `generate`
  when (and only when) invoked via the `generate_episode.py` shim, matching its
  documented contract. The `morning-signal` console script still shows help on a
  bare call, so a stray invocation never silently burns an episode. (ROADMAP P1.6.)

## [0.1.2] — 2026-06-08

### Changed

- Pin `flow-doctor==0.5.0` (was `0.5.0rc3`). flow-doctor cut its first
  stable release after a ~4-week soak with morning-signal as the first
  live consumer; this drops the rc pin onto the final tag now that it's
  on PyPI.

## [0.1.1] — 2026-06-07

First stable release. Seals the `0.1.1rc1`–`rc12` line — no code-behavior
changes from `rc12`; this is the release-discipline cut that makes the package
something an outside self-hoster can depend on. `0.1.x` is deliberate: the
product is real and runs daily, but the CLI + config surface may still evolve,
so this is *not* sold as a finished `1.0`.

### Changed
- **Docs honesty pass for outside users.** README gains an explicit upfront
  **Requirements** block (Python 3.9+, ffmpeg, an Anthropic key always, plus a
  TTS backend = AWS *or* GCP, plus an S3 bucket to publish), and states plainly
  that there is **no fully-offline mode yet** — TTS and publishing both require
  a cloud account today. A local/offline TTS engine and a zero-cloud output
  backend are called out as the most-wanted contributions (the `tts.synthesize()`
  and publish seams are built to make them drop-in). Corrected the stale package
  docstring that still referenced SES email (retired in favor of Telegram).

## [0.1.1rc12] — 2026-05-30

Resync release: PyPI publishing had lapsed after `rc8` (rc9/rc10/rc11 were
version-bumped but never tagged, so never published). This release republishes
the accumulated unreleased work — the prompts-SSM→S3 / public-topics / search-
telemetry block below (the rc11-era content) plus the 2026-05-30 fixes — and
the publish workflow now triggers on push to `main` with `skip-existing`, so the
cadence can no longer silently lapse.

### Fixed
- **TTS speed-adjust resilience (`tts._adjust_speed`).** `ffmpeg atempo` now retries with backoff (3 attempts) and, on final failure, raises loud while logging ffmpeg's captured stderr (previously swallowed by `check=True`). The 2026-05-30 Sat AM edition died with `SIGABRT` when the static ffmpeg build hit a transient allocation failure under memory pressure on the ~916 MB shared host — killing the whole episode at the final step. (#50)
- **Meta-narration no longer leaks into segment audio.** Three-layer fix after the 2026-05-30 edition shipped the model's process talk ("I need to search…", "Based on the search results, I now have… Here's the segment:") to audio on 3/5 segments:
  - `_final_text_after_last_tool` keeps only the text the model writes *after* its last `web_search` — removing pre/inter-search narration positionally (the primary defense), and hardened the per-segment instruction. (#52)
  - `_scrub_segment` now scans every paragraph (not just leading) and drops stray `---` separators (#51); patterns rewritten high-precision so legitimate copy ("Let's start with the numbers.", "Great news from the cosmos.", "search results" as a news topic) is preserved (#53, #54).
  - Companion system-prompt "web search hygiene" rule shipped in alpha-engine-config #387.
- **Circuit-breaker no longer truncates to near-empty.** `enforce_char_budget` collapsed a 2,168-char markets segment to 4 chars when number-dense copy ("26,972.62") left the last sentence boundary very early. It now falls back to a word-boundary cut near the cap, always retaining ~the full budget. (#53)

### Changed
- **Prompt loading migrated from SSM to S3 (production bootstrap path).** `aws._maybe_load_from_ssm` now fetches the three prompts (`prompt.md`, `prompt_weekend.md`, `prompt_public.md`) from `s3://{s3_bucket}/{prompts_s3_prefix}<file>` instead of from `/morning-signal/prompt-md` etc. in SSM. Triggered by `prompt_public.md` crossing the 8,192-char SSM Advanced-tier cap (alpha-engine-config #336 → #338 → #339). SSM keeps the small structured / secret material (config-yaml, anthropic-api-key, telegram creds) where SecureString is the right home; S3 keeps content whose size is a function of the product, not the storage. The morning-signal-runner-role already holds `s3:GetObject` on `morning-signal-podcast/*`, so no IAM change ships in this PR. New required config keys in `config-yaml`: `s3_bucket` (already present for podcast uploads — reused) and optional `prompts_s3_prefix` (default `prompts/`). If `s3_bucket` is missing from config-yaml, boot fails loud with `RuntimeError` per `feedback_no_silent_fails`. Weekend + public prompts remain optional in S3 (rollout window tolerance), with WARN-on-fallback to the weekday prompt for weekend (matches prior SSM-optional semantics). Companion PR: cipher813/alpha-engine-config #339 (sync.sh push to S3, README rewrite, prompt_public.md restored to full 11,353-char version). Old SSM prompt params are not deleted in this PR — they stay as rollback insurance during migration; a follow-up cleanup PR removes them after the soak validates the S3 path.

### Added
- **Public-topics-mode soak substrate (`topic_rotation.py` + extended `load_prompt` + extended `claude.generate_script` + SSM `prompt-public-md` load).** Gated behind new `public_topics_mode` config flag (default `false` — personal-prompt behavior unchanged). When flipped on:
  - `load_prompt(public_mode=True)` loads `prompt_public.md` (the 10-topic catalog shipped via `alpha-engine-config` PR #336) instead of `prompt.md` / `prompt_weekend.md`.
  - `generate_script` computes 5 active topics from `(date_str, edition)` via `topic_rotation.active_topics_for_edition` — 3 fixed (Markets & Economy, Politics, Technology) + 2 rotating wildcards from a 7-set, deterministic round-robin (slot1 period 7, slot2 offset-incrementing). All 21 unordered wildcard pairs are visited once across editions 0–20 (no pair repeats in the first ~10.5 days); each wildcard surfaces exactly 4× across the 14-edition / 7-day soak window.
  - The 5 active topics are injected into the dynamic user message as `Active topics for this edition (cover only these, in this order, ~400 words each): ...`. The system prompt (cached) carries all 10 topic templates; the user message selects which 5 are active.
  - `aws._maybe_load_from_ssm` opportunistically pulls `/morning-signal/prompt-public-md` from SSM (optional — absent param is fine when soak is off). With the flag flipped on but the SSM param missing, `load_prompt` hard-fails loudly at episode generation time (no silent fallback to personal prompt, per `feedback_no_silent_fails`).
  - Epoch date `2026-05-28` anchors the rotation; AM 2026-05-28 = edition index 0 → wildcards (World, Music). Operator soak plan: pair this with `claude_model: claude-haiku-4-5` + `web_search_max_uses: 5` to validate the public-app cost target (~$0.20/edition vs current Sonnet ~$0.50–0.65) end-to-end before building the iOS-app topic-pack-cache infrastructure. (See `topic_rotation.py` docstring + the prompt_public.md README entry in alpha-engine-config.)
- **Per-search `web_search` telemetry (`search_telemetry.py`)** — sibling to `cost_telemetry.py`. Extracts each `server_tool_use` block (the query Claude issued) plus its matching `web_search_tool_result` block (the URLs Anthropic returned), pairing them by `tool_use_id`. Writes one JSONL line per search to `episodes/{date}-{edition}.searches.jsonl`. Called from `claude.generate_script` right after `record_call_cost`. The cost telemetry already captured the *count* of `web_search` requests via `Message.usage.server_tool_use.web_search_requests`; this complements it with the *content* of each search so high-frequency queries and frequently-cited domains can be identified and migrated to curated RSS / direct `web_fetch` sources. Motivation: the 2026-05-27 cost trace showed per-edition Anthropic cost at ~$0.47–$0.66 (vs the stale ROADMAP "~$4.20/mo" estimate), with 58% of that cost going to `cache_create` tokens that originate as `web_search` result content. Reducing search volume — not tightening the cache — is the dominant cost lever, and that requires knowing *what* the model searches.
- **`analyze_searches.py`** repo-root analyzer. Reads every `episodes/*.searches.jsonl` and prints two tables: top-N normalized queries (lowercased, punctuation-stripped — surfaces patterns Claude re-asks across editions) and top-N cited domains (host extracted from result URLs — surfaces direct-fetch candidates). `--episodes-dir` and `--top` flags. Run after a few days of telemetry to pick the first RSS feeds / `web_fetch` substitutions to ship.

## [0.1.1rc10] — 2026-05-25

### Changed
- **Anthropic prompt caching enabled on the production prompt.** The ~1.3K-token static prompt now ships as a `system` block with `cache_control: {"type": "ephemeral"}` (5-min TTL); the user message shrinks to the dynamic preamble (date + edition label). Inside one `messages.create` call, the `web_search` tool loop triggers N+1 inference passes (one per search-decision + final synthesis), each of which re-reads the conversation prefix. Before this change every pass paid the full $3.00/1M input rate on the prompt; now pass #1 pays the $3.75/1M cache-write rate once and passes 2..N pay the $0.30/1M cache-read rate (10× discount). Cross-call (AM→PM) hits don't apply — the 12h gap exceeds the 5-min TTL — but intra-call savings are typically ~80% on the prompt-token portion of input. Lib-side telemetry already captures `cache_read_tokens` / `cache_create_tokens` (lifted in v0.31.0+) so the JSONL records the discount automatically; no `cost_telemetry.py` change required.
- **`web_search` `max_uses` cap added (default 20).** The `web_search_20250305` tool now ships with `max_uses` taken from `config.web_search_max_uses` (default 20). Web search is billed at $10/1k requests by Anthropic; an uncapped tool spec lets a runaway loop or malformed prompt rack up unbounded server-tool fees. 20 sits above the empirical typical (~15 searches per episode for the 9-segment briefing) so this is insurance, not throttling. Config knob is optional — omit it and 20 applies.

## [0.1.1rc9] — 2026-05-25

### Changed
- Bumped `alpha-engine-lib` pin from `>=0.33,<0.34` to `>=0.36,<0.37`. The 0.33→0.36 minor bumps land `pipeline_status` v0.34/v0.35.1/v0.36 + Option-D execution-picker substrate, the `ssm_dispatcher` Python CLI chokepoint (v0.35.0), and an `LLMJudgeReranker` deletion (v0.34.0) — none of those modules are consumed here. The `cost` and `trading_calendar` surfaces morning-signal does consume are byte-identical across the range (verified via `git diff v0.33.0..v0.36.1 -- src/alpha_engine_lib/{cost,trading_calendar}.py` → empty).

## [0.1.1rc8] — 2026-05-25

### Changed
- **`cost_telemetry.record_call_cost` now delegates to `alpha_engine_lib.cost.record_anthropic_call`** (lifted in lib v0.33.0). The local helper shrinks ~50 → ~15 lines; lib owns token + tool-request extraction, default rate-card lookup, USD recompute, and JSONL-ready dict shape. morning-signal still stamps `date` + `edition` onto the record and writes the per-episode JSONL. Public API unchanged: `record_call_cost(*, msg, date_str, edition, episodes_dir) -> float`. JSONL file shape preserved. (PR #28 + alpha-engine-lib #69 — second-recurrence-lift rule, the lib chokepoint now serves data + executor + morning-signal.)
- Bumped `alpha-engine-lib` pin from `>=0.32,<0.33` to `>=0.33,<0.34`.

## [0.1.1rc7] — 2026-05-25

### Fixed
- **PyPI publish path unblocked.** `alpha-engine-lib` is now on PyPI (`0.32.0`), so the `pyproject.toml` dep flips from a `git+https://…@v0.32.0` direct reference to a standard PyPI spec (`alpha-engine-lib>=0.32,<0.33`). rc6 failed to publish with `400 Bad Request` because PyPI rejects published packages whose metadata contains direct-URL deps. The `[tool.hatch.metadata] allow-direct-references = true` opt-in is no longer needed and has been removed. Library code is otherwise unchanged from rc6.

## [0.1.1rc6] — 2026-05-25

### Added
- **Weekend / non-trading-day deep-dive prompt (`prompt_weekend.md`).** On Saturdays, Sundays, and NYSE holidays the AM edition now ships a deeper (~3,000-word) brief focused on frontier models, research papers, AI infrastructure, applied AI products, tech industry moves, and the open-source / dev ecosystem — markets/macro/portfolio segments are replaced with the tech/AI deep-dive content. Selected via `alpha_engine_lib.trading_calendar.is_trading_day`. Loaded from `/morning-signal/prompt-weekend-md` in SSM mode; falls back to the weekday prompt with a WARN if the SSM param is missing.
- **Non-trading-day PM editions are now skipped.** Cron still fires both AM + PM daily, but `episode.main()` exits cleanly (no Claude / Polly / S3, no failure email) when the PM edition lands on a weekend or NYSE holiday. The single weekend AM "weekend edition" replaces both.

### Changed
- **Script opening line is now pinned via an assistant-prefill on `messages.create`.** Every episode now begins with `Welcome to Morning Signal.` (weekday AM), `Welcome to Morning Signal, evening edition.` (weekday PM), or `Welcome to Morning Signal, weekend edition.` (weekend AM). The prefill technique bypasses the `"Great, I now have enough information to compile the episode…"` preamble that Claude was emitting after web-search tool use. Both prompts also carry an explicit "Output format" section forbidding the preamble.

## [0.1.1rc5] — 2026-05-13

### Fixed
- **Progress spinner and stdlib `logging` lines no longer share a terminal line.** rc3 introduced a `rich.progress` spinner but the stdlib `StreamHandler` was emitting `[INFO]` lines via `\n` writes uncoordinated with rich's `\r` cursor redraws, producing visual smush like `0:01:25[12:34:56] Chunk 1/4 done`. Fixed by sharing a single `rich.console.Console` between `Progress` (`episode._CONSOLE`) and a `rich.logging.RichHandler` wired up in `cli._setup_logging()` when running in TTY mode. Non-TTY (systemd journal, cron) keeps the plain stdlib `StreamHandler` unchanged.

## [0.1.1rc4] — 2026-05-13

### Fixed
- **`morning-signal generate --dry-run` errored with `No such option: --dry-run`.** rc3 added the flag at the argparse layer in `episode.main()` but missed wiring it through `cli.generate`'s typer signature. Typer rejected the unknown flag before it reached argparse. Added `dry_run: bool` option on the typer command + argv-translation entry. (+1 test verifies the flag reaches `episode.main()` through the typer wrapper.)

## [0.1.1rc3] — 2026-05-13

### Added
- **`--dry-run` flag on `morning-signal generate`.** Validates setup (config + prompt readability, Anthropic key presence, AWS credential resolution, output paths) without making any Claude / Polly / S3 calls. Useful for post-`init` sanity checks and CI smoke. (+1 test)
- **TTY-aware progress display during generation.** When `morning-signal generate` runs in an interactive terminal, a `rich.progress` spinner shows the current phase (`Generating script (Claude + web search)` → `Synthesizing audio (Polly)` → `Publishing to S3` → `Done`) with elapsed-time counter. Non-TTY contexts (systemd journal, cron output) keep the existing verbose log lines unchanged.
- Final log line now reports total elapsed seconds: `Done in 72s.`

### Changed
- `httpx` and `anthropic` SDK loggers raised from INFO to WARNING. The raw `HTTP Request: POST https://api.anthropic.com/v1/messages "HTTP/1.1 200 OK"` line was visual noise next to the new progress bar.

## [0.1.1rc2] — 2026-05-13

### Fixed
- **`morning-signal generate` failed with `Could not resolve authentication method` when run via the init wizard's smoke test (and any other invocation that relied on `~/.config/morning-signal/.env` for the Anthropic key).** The wizard wrote the key file but the CLI never loaded it; the subprocess inherited a blank env. The CLI now auto-loads `./.env` and `~/.config/morning-signal/.env` at startup (CWD takes precedence). Loading is skipped under `MORNING_SIGNAL_USE_SSM=1` so production hosts continue getting secrets from SSM, and explicit env vars always win over file fallbacks. (+9 tests)

## [0.1.1rc1] — 2026-05-13

Release-candidate cut for internal dogfooding before promoting to 0.1.1 stable.
`pip install morning-signal==0.1.1rc1` or `pip install --pre morning-signal`.

### Added
- 5 bundled prompt presets in `morning_signal/data/`: `generic-news`, `tech-only`, `markets-only`, `local-news`, `blank`. Init wizard offers them.
- CodeQL workflow for weekly automated security scanning.
- Dependabot config: weekly pip + GitHub Actions updates, grouped by AWS / tooling. GH-actions updates bundle into one weekly PR.
- README "Alpha disclaimer" section signaling that the 0.1.x interface may change before 1.0.

### Repo hygiene (not user-visible)
- Branch protection on `main`: status checks required on Python 3.9–3.12, linear history enforced, no force-push, no deletion.
- Tag protection ruleset for `v*.*.*`: tags are immutable (no deletion, no force-update).
- Repo metadata: description, homepage (PyPI), topics set for GitHub profile pin card.

## [0.1.0] — 2026-05-13

First public release on PyPI.

### Added
- `morning-signal generate` — full pipeline: Claude with web search → Polly TTS → ffmpeg speed adjust → S3 upload + RSS feed regen + optional SES success/failure notification.
- `morning-signal preview <prompt-file>` — script-only run with an alternate prompt file for cheap iteration.
- `morning-signal subscribe` — print feed URL + Apple Podcasts / Overcast / Pocket Casts subscribe instructions.
- `morning-signal version` — print package version.
- `morning-signal init` — 8-step interactive setup wizard: AWS credential check, Anthropic key validation, S3 bucket bootstrap (create + public-read policy + CORS), config.yaml + prompt.md write, secret storage at `~/.config/morning-signal/.env`, scheduler installer (launchd / systemd-user / cron auto-detected, DST-aware Pacific calendar), optional smoke test, subscribe instructions.
- Dual-edition support (AM + PM) with 12-hour temporal news window in the prompt — naturally avoids duplicate content across editions without state-passing dedup.
- Production-mode hooks: `MORNING_SIGNAL_RUNNER_ROLE_ARN` (STS AssumeRole at startup), `MORNING_SIGNAL_USE_SSM=1` (fetch config + prompt + Anthropic key from SSM Parameter Store SecureStrings).
- Front-door dedup with `--force` override + back-door upload fix for same-date regenerations.
- SES success + failure notifications (toggleable in config).
- Backward-compat shim at `generate_episode.py` for legacy systemd/launchd units.

### Notes
- Requires Python 3.9+.
- License: MIT.

[Unreleased]: https://github.com/nousergon/morning-signal/compare/v0.1.1rc5...HEAD
[0.1.1rc5]: https://github.com/nousergon/morning-signal/releases/tag/v0.1.1rc5
[0.1.1rc4]: https://github.com/nousergon/morning-signal/releases/tag/v0.1.1rc4
[0.1.1rc3]: https://github.com/nousergon/morning-signal/releases/tag/v0.1.1rc3
[0.1.1rc2]: https://github.com/nousergon/morning-signal/releases/tag/v0.1.1rc2
[0.1.1rc1]: https://github.com/nousergon/morning-signal/releases/tag/v0.1.1rc1
[0.1.0]: https://github.com/nousergon/morning-signal/releases/tag/v0.1.0