Add post-processing to meeting mode and dictation context

[materemias]

Note: Reopened from #265, which was closed when the fork was deleted. This PR contains the same changes, rebased onto current main.

Summary

This PR adds post-processing support to two areas that previously lacked it:

1. Meeting mode post-processing (new)

Meeting mode previously skipped post-processing entirely. Now, if [output.post_process] is configured, each transcribed chunk is piped through the post-processing command before being stored in the transcript. Previous chunk text is passed as context so the LLM can maintain continuity across 30-second chunk boundaries.

Context is tracked per audio source (mic vs loopback) to prevent speaker bleed between channels. Context is cleared when a meeting stops so it doesn't leak into the next meeting.

2. Dictation mode context (new)

Regular push-to-talk dictation now passes the previous dictation's text as context to the post-processing command, if the previous dictation was within 60 seconds. This helps the LLM produce more coherent cleanup when dictating in quick succession.

The 60-second retention window is intentionally not configurable to avoid config fatigue. It covers rapid-fire dictation without leaking stale context.

Context delivery mechanism

Context is passed via the VOXTYPE_CONTEXT environment variable, keeping stdin as plain current text. This means existing post-processing scripts work unchanged. Scripts that want context can optionally read $VOXTYPE_CONTEXT.

The env var is explicitly cleared before each command invocation to prevent inheriting stale context from the parent environment.

Other changes

• process() now delegates to process_with_context(text, None), removing ~20 lines of duplicated error-handling logic
• Post-processing tracing logs now include changed: true/false to indicate whether the command modified the text
• Example scripts (gemini, ollama, openai) updated to optionally use $VOXTYPE_CONTEXT
• Docs updated (CONFIGURATION.md, USER_MANUAL.md)

Test plan

• Configure [output.post_process] and run a meeting, verify exported transcript shows post-processed text
• Verify post-processing failure falls back to original text
• Verify mic and loopback chunks get independent context (no speaker bleed)
• Verify context clears on meeting stop and doesn't leak to next meeting
• Verify regular dictation gets context from previous dictation within 60s
• Verify dictation context expires after 60s
• Verify existing scripts work without modification (stdin is plain text)
• Verify VOXTYPE_CONTEXT is not inherited from parent environment

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR adds optional context-aware post-processing support to meeting mode chunk transcription and to push-to-talk dictation, enabling external cleanup commands (e.g., LLM scripts) to maintain continuity across chunks/dictations via VOXTYPE_CONTEXT.

Changes:

• Add PostProcessor::process_with_context that passes prior text via VOXTYPE_CONTEXT while keeping stdin as current text.
• Apply post-processing to meeting-mode segments with per-audio-source context tracking and context reset on meeting stop.
• Track recent dictation output for up to 60 seconds and pass it as context into post-processing; update example scripts + docs to optionally use VOXTYPE_CONTEXT.

Reviewed changes

Copilot reviewed 9 out of 9 changed files in this pull request and generated 4 comments.

Show a summary per file

File	Description
src/output/post_process.rs	Introduces context-aware post-processing via env var and refactors process() to delegate.
src/meeting/mod.rs	Adds meeting-mode post-processing and per-source context tracking using a HashMap.
src/meeting/data.rs	Derives Hash for AudioSource so it can be used as a HashMap key.
src/daemon.rs	Adds recent dictation context tracking (60s) and passes it into post-processing; extends tracing logs.
examples/openai-cleanup.sh	Updates prompt construction to optionally include prior dictation context.
examples/ollama-cleanup.sh	Updates prompt construction to optionally include prior dictation context.
examples/gemini-cleanup.sh	Updates prompt construction to optionally include prior dictation context.
docs/USER_MANUAL.md	Documents VOXTYPE_CONTEXT behavior and meeting-mode per-source context.
docs/CONFIGURATION.md	Documents VOXTYPE_CONTEXT behavior and references updated examples.

[Copilot]

context borrows from self.last_chunk_text and is then held across .await calls to process_with_context. Because the method later mutates self.last_chunk_text (via insert), this is very likely to fail to compile with a borrow-across-await error. Consider copying the context into an owned String (or cloning the map value) before awaiting, and/or updating a local context variable as you process segments, then write back to the map after the awaits complete.

[materemias]

Fixed in 4637653. Cloned the context into an owned String before the async loop, using .as_deref() at the call site. The original code did compile (NLL drops the borrow before the insert), but the clone is cheap and avoids fragility if the loop body changes later.

[Copilot]

recent_context is a &str borrowed from self.last_dictation and is passed into async post-processing calls (.await). The function then updates self.last_dictation afterwards, which is likely to cause a borrow-checker error (borrow of self held across await, then mutated). To avoid this, materialize context into a local owned value (e.g., clone the string into an Option<String> and use .as_deref() when calling process_with_context).

Suggested change

	let recent_context = self.last_dictation.as_ref().and_then(|(text, when)| {
	if when.elapsed() < Duration::from_secs(60) {
	Some(text.as_str())
	} else {
	None
	}
	});
	let recent_context_owned =
	self.last_dictation.as_ref().and_then(|(text, when)| {
	if when.elapsed() < Duration::from_secs(60) {
	Some(text.clone())
	} else {
	None
	}
	});
	let recent_context = recent_context_owned.as_deref();

[materemias]

Fixed in 4637653. Cloned the context string into an owned Option<String> and use .as_deref() at all three call sites. Same reasoning as the meeting/mod.rs fix: the original compiled fine under NLL, but the owned value is more robust.

[Copilot]

Logging recent_context at info level can leak the previous dictation text into logs (potentially sensitive/PII) in addition to the current transcription. Consider logging only whether context was present (or its length), and/or moving the context value to debug tracing.

[materemias]

Good catch. Fixed in 4637653. The info level now only logs whether context is present (has_context: true/false) and whether the text changed. Actual text content (input, context, result) is now at debug level only.

[Copilot]

New VOXTYPE_CONTEXT behavior isn’t covered by tests. Since this file already has async tests, consider adding coverage to assert (1) when context is None, the child does not see VOXTYPE_CONTEXT even if it exists in the parent env, and (2) when Some, the child receives the exact context value while stdin still only contains current text.

[materemias]

Added three tests in 4637653:

• test_context_passed_via_env_var: verifies VOXTYPE_CONTEXT is set and stdin contains only current text
• test_no_context_env_var_when_none: verifies the env var is absent when context is None
• test_context_env_not_inherited_from_parent: verifies stale parent env is cleared when context is None