- bound _parse_responses scans to the output_text part's object and to
the content array, so a later sibling item's "text" (e.g. a reasoning
summary) can never be returned as the assistant reply
- do not retry the other flavour on a curl timeout: the request likely
reached the server and may still be generating, so a re-POST of the
conversation could trigger a second billed generation
- remember an unparseable 2xx from the first AUTO attempt and surface
stashed first-attempt errors in both error paths, instead of showing
only the final attempt's transport or HTTP error
- recognize Ollama's model-not-found wording in _names_model so a model
typo is not misread as a missing endpoint
- drop the dead, racy provider-lookup fallback in the generic request
thread: a missing provider ref is a caller bug and now fails loudly
- fix ai_providers_lists_defaults to expect the header the command
actually prints ("Configured providers:"); the test was broken since
its introduction but CI never ran it
- add functional test group 5 (AI command surface) to FUNC_TEST_GROUPS
so the CI parallel target runs it; proftest.c port ranges already
account for five groups
Parser:
- Never return reasoning text as the reply: the no-output_text path
anchors strictly on the message "content" array; a reasoning-only body
(truncated by max_output_tokens, carries only "summary") now fails the
parse instead of leaking the chain-of-thought. Regression test added.
- _find_json_field skips field-name occurrences that are string values
(no ':' after), so {"type":"text","text":"hi"} parses instead of
failing on the value of "type".
- _same_json_object tracks brace depth, so a nested object between the
"text" key and the "output_text" tag (e.g. "annotations") no longer
rejects the part's own text.
- Chat-completions extraction is bounded to the choices array (new
_json_array_end helper), so "message"/"content" in a sibling object
(e.g. a top-level "warning") is not returned as the reply.
AUTO fallback:
- Model errors are classified only from a structured JSON error
envelope; a route-missing 404 page merely mentioning models (e.g.
listing /v1/models) no longer suppresses the fallback. Trade-off: a
model error in a non-JSON body costs one extra fallback attempt.
- Fall back on a curl-level failure: an endpoint that accepts the
connection but never responds no longer burns the 60s timeout without
trying the flavour that works.
- Fall back on an unparseable 2xx: a gateway answering unknown paths
with a 200 error page previously reset the hint and re-probed the same
dead flavour on every request. The body is parsed inside the attempt
loop and no longer re-parsed after it.
- The second attempt's 401/403/429/5xx now wins over the stashed 400/422
payload rejection, so an invalid key or rate limit is not hidden
behind a payload-shape error; a second 404 still defers to the first.
- When the first flavour's endpoint is missing but the other answers
(even unsuccessfully, e.g. 401), the resolved hint moves to the
surviving flavour, so the dead endpoint is not probed first forever.
Settings:
- Reserved payload keys are rejected only when setting a value; removal
is allowed again, so a setting persisted before its key became
reserved (e.g. "input") is no longer stuck in the config.
The setting-name branch for /ai set custom never fired: it passed the
prefix "/ai set custom " with a trailing space to
autocomplete_param_with_ac(), which appends a space itself, so the
matched prefix contained a double space no real input ever has. Even
without that, the primitive completes the first token after the command
while the setting name is the second, and the num_args guard only held
before the name was started. The suggestion list (tools/search/memory/
plugins) predated the /ai set custom backend and matched no real
payload parameter.
Replace the branch with token-position completion (arg 5 via
autocomplete_param_no_with_func) fed by a list rebuilt on demand from
common payload parameters plus the keys already set on the provider,
exposed through the new ai_get_provider_setting_keys() (snapshot taken
under settings_lock). Drop the now-unused parse_args() call whose NULL
result was fed to g_strv_length() unguarded, spamming a glib CRITICAL
on TAB at /ai with zero or five-plus arguments.
Drive both OpenAI-compatible flavours from one code path. Requests
default to /v1/responses and fall back once to /v1/chat/completions when
the provider reports the endpoint missing (404/405/501) or rejects the
payload shape (400/422); the working flavour is cached for the rest of
the run as a first-attempt hint, with the other flavour kept as a
fallback so a backend change self-corrects in-request. A per-provider
override is available via
/ai set api-type <provider> responses|chat-completions|auto
and persisted as api_type= in the provider section.
One payload builder serves both flavours (messages/stream vs
input/stream/store); history and custom settings are serialized once per
request and the per-flavour envelope is assembled per attempt under a
single lock acquisition, so the fallback retry cannot double-count the
prompt. ai_parse_response_typed() dispatches on the request flavour and
splits extraction per envelope: chat completions anchored on
choices[].message.content, responses on the output_text part with a
string-aware backward scan bounded to that part's own object so a
reasoning summary or a truncated body is never returned as the reply.
Endpoint detection classifies wrong-model errors from the structured
error code/type/message when present, so route-missing 404s from
gateways no longer suppress the fallback; it preserves and surfaces the
first payload-rejection error when the fallback also fails, and re-probes
on an unparseable 2xx. resolved_api_type is written only through a locked
helper guarded by a URL/api-type epoch, and api_url is snapshotted under
settings_lock in the request and models-fetch threads, closing
use-after-free and stale-cache races against /ai set provider. The
models-fetch provider ref is taken on the main thread and released on
every worker exit path. Reserved custom-setting keys are extended with
input (store stays writable as a legitimate chat-completions parameter).
Custom setting values were JSON-escaped but emitted unquoted, so any
non-numeric value (e.g. "high") produced an invalid JSON payload and
failed the whole request; quoting the value manually could not work
either, since the escaper turns quotes into \". Emit RFC 8259 scalars
(number/true/false/null) bare and everything else as a quoted JSON
string.
Reserved payload keys (model, messages, stream) would duplicate the
fixed fields with parser-dependent precedence; reject them in
ai_set_provider_setting (surfaced as an error by /ai set custom) and
skip them at payload-build time for settings loaded from hand-edited
prefs.
provider->settings was mutated on the main thread while the request
thread iterates it when building the payload; guard both sides with a
new per-provider settings_lock.
Also refresh stale docs: ai_parse_response no longer mentions the
dropped Perplexity "text" path, the payload docstring says "messages"
instead of "input", and the /ai help example uses a realistic scalar
setting.
Update request endpoint to /v1/chat/completions and switch payload key
from input to messages. Remove store flag and legacy Perplexity response
parsing to standardize on OpenAI's content extraction.
Modify _build_json_payload_from_list to accept an AIProvider parameter
and dynamically merge its custom settings into the JSON payload. The
settings are serialized as additional key-value pairs alongside the
standard model and input fields, enabling per-provider configuration
options without hardcoding them.
Add helper functions to parse hex digits and encode UTF-8 characters. Update buffer allocation to account for UTF-8 expansion and implement full surrogate pair support for characters outside the BMP. Previously, \uXXXX sequences were passed through verbatim; they are now correctly decoded into proper UTF-8 strings.
Add missing `ai_session_unref` calls to prevent memory leaks.
- In `_ai_request_thread`, release session on error paths and
after successful processing.
- In `cmd_ai_start`, release reference after passing ownership
to the AI window.
Replace manual g_free calls with auto_gchar for local_provider_name,
local_model, local_api_key, and response_data to ensure automatic
cleanup and prevent memory leaks.
Add an AI client module that integrates with OpenAI-compatible API
providers (OpenAI, Perplexity, and custom endpoints) to provide
AI-assisted chat within CProof. Users can start sessions with /ai start,
send prompts, receive responses in a dedicated AI window, switch between
providers and models, and manage API keys — all with tab-completion.
Providers are configured via /ai set commands with per-provider API keys,
endpoints, default models, and custom settings. Two default providers
(openai, perplexity) are seeded on first use. Provider state persists in
[ai/<name>] sections of the preferences keyfile with automatic migration
from the previous flat-key format.
The /ai command integrates into the existing command system with 8
subcommands covering provider management, session lifecycle, model
fetching, and conversation clearing. Autocomplete uses the standard
flat prefix-matching chain for reliable tab-completion at every nesting
level. A new ProfAiWin window type is added to the window system.
Architecture:
Async design: HTTP requests run on a background thread (pthread) to avoid blocking the ncurses UI loop; results are displayed on the main thread via direct function calls
Thread safety: AIProvider and AISession use atomic ref-counting and mutex-protected session state; the request thread snapshots all session data before making the HTTP call
Window validation: wins_ai_exists() prevents use-after-free when the user closes the AI window during an in-flight HTTP request (~60s)
Privacy: store:false is sent with every request to prevent providers from persisting conversations or using them for training
Response size limit: 10MB cap with immediate curl abort via CURL_WRITEFUNC_ERROR to prevent OOM
JSON parsing uses unified helpers for both chat responses and error
envelopes with consistent escape decoding. The response parser tries
Perplexity /v1/responses "text" field first, then falls back to OpenAI
"content". Error parsing extracts provider error.message from the
standard envelope format. Model parsing handles multiple API response
formats (OpenAI list, Perplexity, array) including edge cases.
Tests include 470+ lines of unit tests covering provider management,
session lifecycle, JSON parsing (multiple formats), autocomplete cycling,
and error handling, plus functional tests for /ai command dispatch.
A stub_ai.c module isolates unit tests from UI dependencies.
